Slot machines are (purportedly) random dial-spinners that stop in ultimately random ways (that are adjusted for particular pre-set overall payoff levels, but not individual spins). In this sense, the slot machine is nearly a random device, but even the computer-based random number generator of modern slot machines is not 100% random and, in a sense, every spin could be predicted at least in principle.So, as far as anybody can tell in practice, each flip or each jerk of the one-armed bandit, is random. We still can say much about the results: We can't predict a given coin flip or slot-pull, but we can predict the overall net result of many pulls, to within some limits based on statistical probability theory--though never perfectly.
On the other hand, a casino is a collection of numerous devices (roulette wheels, poker tables, slot machines, and so on). Each is of the same probabilistic kind. Nobody would claim that the take of a casino was not related to these devices, not even those who believe that each one is inherently probabilistic. To think that would be to argue that something other than physical factors made up a casino.
But the take of a casino on any given day cannot be predicted from an enumeration of its devices! The daily take is the result of how much use was made of each device, of the decision-making behavior of the players, of the particular players that were there that day, of how much they were willing to lose, and so on. The daily take is an 'emergent' property of the assembled items. Interestingly, nonetheless, the pattern of daily takes can be predicted at least within some limits. This is the mysterious connection between full predictability and emergence, and it is a central fact of the life sciences.
Genes exist and they do things. On average, we can assess what a gene does. Clearly genes underlie what a person is and does. But each gene's net impact on some trait depends not just on itself, but on the rest of the genes in the same person's genome, and countless other factors. A particular individual's particular action is simply not predictable with precision from its genome (or, for that matter, its genome and measured environmental factors). There are simply too many factors and we can't assess their individual action in individual cases, except within what are usually very broad limits.
Brain games
A common current application of the issues here is to be found in neurosciences. There is a firm if not fervid belief that if we enumerate everything about genes and brains we'll be able to show that, yes, you're just a chemical automaton. Forget about the delusion of free will!
 |
| Location of the amygdala; Wikipedia |
Day after day, in the media and in the science journals themselves, the promise is made of ultimate (often, of imminent) predictability even of complex emergent phenomena, from examination of their parts. If we just have enough sequencers, fMRI machines, and other kinds of technology, everything will work out. Not to worry!
So the Human Connectome Project, exploits the 'omics idea that if we mindlessly enumerate every single little thing we can understand every single big thing, is funded and off and enumerating.....every connection between every neuron in the brain (starting, we think, with 'the' mouse, whatever that means). Mindlessly is the right word, because the investigators of such things often proudly proclaim that they are not testing any hypothesis about Nature: this is pure Baconian empiricism, something we've discussed in earlier posts: collect all the facts and the theory will emerge automatically. There seems to be a feeling of imminent triumph that, like the priests of old, we The Scientist will be able to see inside your very soul to see what you are really like, no matter how much you may delude yourself that you are a free agent.
Clear-cut cases of prediction in complex systems from specific identified elements do exist, due to individually very strong factors. They are usually rare, but they addict us to the idea that all cases--all behaviors or even all thoughts, will be predictable by enumerating all causal factors and their effects. But this is, at best, not practicable. Is it an ultimate illusion?
So why the persistent belief to the contrary?
Could it be that really, truly, and ultimately when so many countless probabilistic factors interact to generate a net result, our ability to predict them other than in a few special cases is inherently limited? Could it be that our claims to do otherwise are, in fact, no more than a current version of Delphic mumbo-jumbo that has always existed in society? Whether or not that is true, science, like religion, is not likely to agree to that.
Why is there such reluctance to simply accept limits to our knowledge, or perhaps even to our ability to know things by applying current methods? Is it just arrogance, careerism, profit-chasing? Is it ignorance of the landscape?
One thing is that of course we cannot apply scientific methods that we haven't yet discovered. There are programs and even organizations, like the Santa Fe Institute of which Ken is an external faculty member, that are dedicated to working out an understanding of complexity. We think it's fair to say that they haven't solved the problem!
At present, a nay-sayer may be viewed as someone who is anti-science, or perhaps even being mystical. After all, either things are material or they aren't! If they are material, should we not be able to understand them? If they are numerous or individually small, doesn't the history of science show that instrumentation and technology need to be brought to bear on the problem?
The answer to these questions is certainly 'yes'. We're not mystics. But physical problems need not be amenable to the kinds of solutions we currently have, any more than astrology solved problems when observing the stars and planets was the technology of the time. Our society certainly believes in technology and even more so, perhaps, in the idea that technology is for making a profit. That's the often explicitly stated that the point of science is its application, that we do this for our careers and labs, or for patients, or for society at large.
But it is not defeatism to ask whether the current approaches, based on 400 year-old Enlightenment-derived methods and concepts, are obsolete for the kinds of questions we are now asking (no matter how powerful they were for lesser questions that were successfully answered). It could potentially help to withdraw resources from business as usual as a way of trying to force more creative thinking--but there's no guarantee that, if, or when, it would work to stimulate the next Darwin or Einstein.
It is similarly not out of line to ask, as regular readers know we ask regularly, whether much of what is being supported in science is on wrong trails, even if good for maintaining funding and other sorts of momentum, by diverting funds from things more likely soluble with traditional approaches, like diseases that really are genetic and for which genetic treatments would be fantastic.
And it is not out of line to ask whether when there are so many really serious human ills in the world, that have nothing to do with genes (or, for that matter, with science), that resources are often wrongly being used to maintain an academic welfare system, the way passing the plate maintains religious establishments on the promise of Things to Come.
As we have often also said often, triggered by yet more grandiose claims in the news or journals, complexity due to multiple interacting but individually small factors is the challenge of the day. It is even more challenging to the extent that really, or for all practical purposes, these factors are probabilistic results of large numbers of interacting, individually minor, factors.
If that's the case, we are back in the 1800s, when it was discovered that every year a predictable number of people will commit suicide, and by predictable arrays of methods, but yet this can rarely be predicted for individuals. That kind of problem was perhaps first recognized more than a century ago, but is still with us.
And it's a no-brainer to recognize that.
I suspect my position may be judged irrelevant, but in many areas we do a great job of postdiction, meteorology as the principal example. That is, we can do a wonderful job of explaining why something happened post hoc, even though we could NOT have predicted the same event. And, I would argue, it is postdiction in general that defines most of what we refer to as "science" prediction. To demand real-world prediction is to adhere slavishly to some logical positivist idea of science that has never amounted to much. Yes, of course, we should "predict" the results of our ceteris paribus experiments designed precisely to test specific causal hypotheses---something we do quite well, from experimental physics to even in the social sciences. But that has nothing to do with real world predictions, nor should it. It is this idea that what we learn should have real-world predictive consequences that is at fault, and, I believe, is the major factor behind your concerns.
ReplyDeleteMe? I am happy when we can postdict, that is, explain why something happened in the real world (e.g., plane crashes, earth-quakes). For the rest, we run ceteris paribus, closed-world experiments, often resembling nothing in the real-world. And that is a good thing: general prediction is NOT the test of good science, and never has been.
Well, I see your point but I think this time I don't really agree, at least in several senses.
DeleteFirst, predictive power is routinely described as what theory does after it has 'retrodicted' many observations. From its inception in the Enlightenment, I think that has been a rather standard tenet of philosophy of science and 'the scientific method'.
Second, retrofitting can often be done in many different ways by adjusting the relevant theory being used. The way to tell if it was done right is replication. That's a form of prediction, in a sense, because you confirm the idea when you get the expected--predicted replicating--result.
In chemistry and much of physics, without indulging in physics-envy, things are entirely predictive, in that artillery and spaceflight and electronics and so on work because we can predict what will happen. There may be error, of course, and even if it's not all due to measurement alone, but the accuracy of prediction to error is much higher and far better than guessing, which is too often the case in the 'softer' fields of science.
In much of biology it's the same. As examples, I can predict DNA sequences and PCR amplification and cloning, and gene expression, making transgenic mice, and many other aspects of experimental biology.
To me it is in the social, epidemiological, evolutionary and other behavioral observational fields where I think prediction fails very badly compared to these other fields.
Since these observational fields are presumably dealing with the same physical universe of causation, I think there must be some better way to understand what's going on--or to prove conclusively that prediction cannot be improved beyond some estimable point because (for example) too many factors are interacting, even if each is faithfully following natural law.
I do see why it is satisfying if you can fit experience to some explanatory framework, but how can you tell if you are really accounting for it in causal terms rather than Just-So plausibility story-telling?
As to meteorology, I have fond feelings as I was a meteorologist once, and at least where we live they are vastly better today than they were. I am not sure how much of that prediction is based on fitting empirical present data to past pattern (that is, not heavily guided by theory) of countless past events to generate an expected near-term future--and to what extent physics and hydrodynamics is really being used in a specific rather than very general sense. I'm just out of date. Maybe when I retire, soon hopefully!, I will do some investigation of the state of play there these days.....